Vehicular emissions are a concern of automobile industries and oil companies due to their impact on both human health and global warming. Less pollutant technologies and more efficient fuels have been developed in the last years driven by constrains imposed by government regulations. However, the estimation of such improvements in the real scenario is very hard to be evaluated due to many reasons mainly because the difficulty of reliable emission models. The main oil companies and automakers usually perform emissions tests to support the development of new fuels and to evaluate new production technologies. So, a large amount of data is generated that can be useful to develop data based models through data mining techniques. In this work, a data base that has been collected over more than 10 years is used to build neural network models for pollutant emissions, given the gasoline properties and vehicle characteristics. The resulting models have achieved performances comparable to the uncertainties inherent into the emissions tests. The models are used to estimate the emissions impacts of the new gasoline composition that should be produced in Brazil in a near future. The results show that the new fuel composition allows the reduction of the main pollutants concentrations without a significant drop in vehicle fuel economy.